The Agent Statistics Everyone Quotes Are a Year Old

A snapshot as at 19 July 2026.

I went looking for what the evidence base on enterprise agent adoption looked like in July 2026. What I found was more interesting than what I was looking for: almost every widely quoted statistic about agentic AI outcomes is a year old, and is being recirculated with fresh datelines.

The provenance of the numbers you have seen

Three figures dominate almost every article about agent adoption. Here is where each actually comes from.

• "95% of AI pilots show no measurable P&L impact." This is the MIT NANDA study — published July 2025. It is now two years old in AI-industry time, which is a meaningful interval.
• "40% of agentic AI projects will be cancelled by 2027." This is a Gartner prediction from June 2025. Note what it is: a forecast about a future year, made over a year ago. It is not an observation of what has happened. It is regularly quoted as though it were a measured outcome.
• "88% of pilots never reach production." This one is attributed variously to Anaconda and Forrester across many sites. I could not locate a primary report behind it. I am not asserting it is wrong — I am reporting that I could not find its source, which is itself a reason not to put it in a board pack.

The same effect shows up in commentary. The most prominent sceptical piece surfacing in searches for July 2026 — an op-ed on agents being ambitious, overhyped and still in training — was originally published on 3 January 2026 and has been syndicated continuously since. It reads as current. It is six months old.

Why this is worse than ordinary staleness

Old data is not automatically bad data. The problem here is specific: the interval between when these numbers were collected and now contains most of the change they are being used to reason about.

Consider what has shipped since mid-2025. MCP has gone from a young protocol to a Linux Foundation project with a registry of nearly ten thousand servers and an enterprise authorisation extension. Coding agents moved from synchronous chat assistants to background and remote execution. Agent frameworks from Microsoft, Google, OpenAI and Anthropic reached general availability. Model capability on agentic coding benchmarks moved substantially and is now visibly saturating near the top of the range.

A pilot failure rate measured before any of that existed tells you about a different technology stack than the one you are being asked to fund. Quoting it as current evidence is not conservatism. It is a category error dressed as prudence — and it can support a bad decision in either direction, either killing a viable programme or, just as easily, reassuring someone that everyone else is failing so a poor result is acceptable.

What I could not find, which is itself the finding

Searching the three weeks to 19 July 2026, I found no significant new enterprise adoption study, no new ROI research, and no landmark agent benchmark result. Leaderboards refreshed. No new frontier number landed. There was no new sceptical analysis of substance either — the critical case is being made with the same year-old material as the optimistic one.

What I did find in that window was shipping: framework releases, changelogs, security disclosures. Those are reliable evidence of what was built. They are not evidence of what worked. A changelog tells you a vendor shipped background subagents; it tells you nothing about whether any organisation got value from them.

So the honest summary of the current evidence base is: we have excellent data on agent capability and almost no current data on agent outcomes. That asymmetry should make anyone quoting an outcome statistic — in either direction — considerably more careful.

The practical response

If the public evidence base is stale, the answer is not to find a better statistic. It is to stop outsourcing the question.

• Measure your own baseline before you deploy. Cycle time, defect escape rate, review burden, cost per task, whatever the agent is meant to improve. Without a pre-deployment number you will be arguing about vibes in six months.
• Define the failure condition in advance. "40% get cancelled" is only frightening if you have no criteria of your own. Decide up front what result would cause you to stop, and at what date you will check.
• Instrument the loop, not the demo. Agent value shows up in aggregate over many runs — token cost, retry rates, human intervention frequency, task completion without escalation. None of that is visible in a successful demonstration.
• Treat vendor productivity claims as hypotheses. Percentage improvements in press releases are supplied by parties with an interest in the number. They are a reason to run a trial, not a substitute for running one.
• Date every statistic you repeat. If a number appears in an internal document without a date and a source, it should not survive review. This single habit would eliminate most of the problem described in this post.

A note on the sceptics and the enthusiasts

It is worth noticing that both camps are drawing from the same exhausted well. The enthusiast cites vendor benchmark gains; the sceptic cites the 2025 pilot-failure figures. Neither is describing what enterprise agent deployment looks like in the middle of 2026, because — as far as I can establish — nobody has published a serious current study of it.

That gap will close. Someone will do the work, and when they do it will be worth reading carefully, including the methodology and sample size. Until then, the most defensible position on agent ROI is the least satisfying one: the public evidence is thin and dated, so measure your own and report it honestly.

We have excellent data on what agents can do and almost no current data on whether they are working. Anyone quoting a confident outcome number in mid-2026 is quoting something from 2025.

Sources and provenance

• MIT NANDA study on AI pilot P&L impact — July 2025 (origin of the "95%" figure)
• Gartner prediction on agentic AI project cancellation by 2027 — June 2025 (a forecast, not an observation)
• "88% never reach production" — attributed to Anaconda/Forrester by numerous secondary sites; primary source not located
• "Meet the AI agents of 2026 — Ambitious, Overhyped and Still in Training" — originally published 3 January 2026, widely syndicated since
• OWASP GenAI Security Project, State of Agentic AI Security and Governance v2.01 — 11 June 2026, one of the few genuinely current datasets, though scoped to security rather than ROI